GNU bug report logs - #63088
[PATCH 0/3] Add Lc0

Previous Next

Full log

View this message in rfc822 format

From: zamfofex <zamfofex <at> twdb.moe>
To: 宋文武 <iyzsong <at> envs.net>
Cc: 63088 <at> debbugs.gnu.org
Subject: [bug#63088] [PATCH 0/3] Add Lc0
Date: Thu, 11 May 2023 18:25:33 -0300 (BRT)

> Will we able to train a model from source?  And how much will it cost?
> 
> As far as I know, guix haven't decided to accept pre-trained models:
> 
> https://yhetil.org/guix/xanfHBZT3lYlyrr_OqHHMWkunLeZZlcxzY37_T3TeZo7mfJClD5-OTbkXDH2f3lMTkn94YIFVUj-Z31BP2Wj0W2rISNP6glC2PzXcPdb560=@protonmail.com/

I don’t think it is feasible. Lc0’s approach is based on self‐play reinforcement‐learning, which in effect means that training starts with no knowledge about chess except for its rules (playing seemingly random moves), and going from there to learn increasingly more about it. So, in practice, this means it would take multiple months or at least several weeks to attain a neural network model that is anywhere close to as effective as the ones provided and pre‐trained. Besides, I believe training requires a (reasonably decent) GPU.

But the network isn’t prohibitively large and its license is not proprietary. I’ll also note that Stockfish (packaged immediately above my Lc0 package) *does* itself include a provided pre‐trained network and no‐one has really complained about it, so I think it should be fine here too.

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.